


Backed by @galaxyhq, @vaneck_us, Crunch's network of 10,000+ ML engineers delivers production-ready ML models for global scale.


Every dataset is a question nobody has asked yet. Institutions hold two things in surplus: data they have not fully used and problems they have not fully solved. Put them together and you get a research challenge. Leave them apart and you get storage costs.

Most networks promise that everyone can earn. Crunch makes a narrower promise: anyone can propose a solution to a problem, and the best ones are rewarded.

A bank can get models from outside researchers without showing them its raw data. This works in 2 ways: 1. Obfuscated data. DataCrunch gives researchers 1,150 anonymized features to predict returns on the 3,000 most liquid US stocks. They never see what the features are. 2. Private data. The data stays in the institution’s own cloud, researchers submit models, the models run in a trusted execution environment, and each one is scored against held-out outcomes. Either way, the institution keeps its data and still gets the research.

Structural breaks are hard to detect. This one has a release date.

Watch the full episode: Minds & Machines EP2 with @herelle_jean.
Horst Simon is one of the people who shaped supercomputing, the field today's AI runs on. His forty years of work span Boeing, NASA and Berkeley Lab, and include two Gordon Bell Prizes and the TOP500 ranking of the world's fastest computers. In Episode 2 of Minds & Machines, he talks with Jean Herelle (@herelle_jean) about AI, supercomputing and the research that ADIA Lab, which he directs, has run on Crunch since 2023.

A gigawatt data center uses ten million times the power of a human brain. Horst Simon says that is the sign something is wrong. The two-time Gordon Bell Prize winner spent forty years building faster computers, and he is not convinced that bigger models will get us to AGI.

Research on Crunch has a multiplayer mode. We think more people should use it.

One of cancer research’s fastest growing methods is up 13x since 2021, yet fewer than 1% of its papers come from lower-income countries due to high costs. Kalin Nonchev’s (@nonchevk) DeepSpot tackles this gap by predicting gene maps from standard tissue slides available in labs worldwide.


21,000 models had to spot the moment a time series changes its behaviour, with the data arriving in real time. ADIA Lab set the problem. 1,700 researchers worked on it on Crunch for 5 months. The winning model scored 69.35% AUC. A random guess scores 50%. At the @adia_lab Symposium in Abu Dhabi, Jean Herelle walked through the 10 winning methods and the top researchers behind them. youtube.com/watch?v=lad_PKvp…



Full report here: newsletter.semianalysis.com/…

AI models score higher and higher on benchmarks. GDPx tests whether they can do real specialist work, with 206 expert-level tasks across 11 professional fields. The top frontier model scores just 12%. SemiAnalysis published GDPx and built it with 14 specialist partners in fields such as biology, finance and chip design. Models get up to 8 hours and access to real tools. In finance, almost every model scores zero. The hardest tasks still need deep domain expertise and human judgment. At Crunch, we combine both. Our researchers build the models, and we judge their work on real-world performance.

His models did the talking. The best model for predicting how fat cells respond to a gene switch-off came from a self-taught data scientist. Laurent Renard is a signal processing engineer by training and an actuarial Python developer at BNP Paribas Cardif. In Part 1 of the Broad Institute’s Obesity ML Competition, his model finished first, more than 50% ahead of second place. He also ended 2025 at number one on a DataCrunch leaderboard, Crunch’s weekly ranking of US stocks. In only two years he has gone from 848th to 17th of all time, collecting two gold medals and the Crunch Grandmaster title along the way.